[Torrero]

Motivation

Exporting large quantized models to Circle (e.g. Qwen3-VL-4B, whose vision subgraph unrolls attention per chunk×head×block into ~10^5–10^6 operators) takes hours of single-threaded CPU with no artifact.

py-spy profiling showed ~93% of export time is spent in TICO's own pass machinery, and ~7% for torch.export:

Phase	Share before
const propagation	~30%
fake-quantize decomposition	~29%
Circle O1 rule matching	~26%

Root causes

1. CircleGraph._operator() rebuilt as_list(subgraph.operators) — an O(N) flatbuffer→list copy — on every accessor call, making every graph traversal O(N²).
2. operator_is_live() ran a full-graph BFS (backward_operators) per operator per rule match → O(N³) in the rules engine.
3. get_constant_from_tensor() re-walked scale/zero-point constant chains without memoization, once per fake-quantize node.

Changes

• tico/circle/graph.py
  • Materialize _operators_list / _tensors_list once at construction (instances are revision-pinned immutable snapshots — the session rebuilds them on any subgraph mutation, so caches cannot go stale).
  • Memoize operator_inputs / operator_outputs.
  • Add live_operators(): one BFS per graph, cached as frozenset (immutable — callers cannot corrupt the shared cache).
• tico/circle/passes/optimization/_pattern_utils.py
  • operator_is_live() uses the cached live_operators().
• tico/passes/decompose_fake_quantize_tensor_qparams.py
  • Memoizing wrapper for get_constant_from_tensor() with a per-pass-run cache keyed by node name.
• Tests: 5 new unit tests covering accessor memoization, snapshot stability, liveness caching/immutability, dead-branch reporting, and memo reuse.

Correctness

• No semantic changes: read-path caching of pure functions only.
• Exported artifacts are byte-identical (sha256-verified on quantized Qwen3-VL-4B decoder layers).
• tico-circle verify passes on all produced artifacts.
• Cache invalidation follows the pre-existing revision scheme (CircleOptimizationSession.graph()); the FX-pass cache lives exactly one call() and reads only immutable constant chains.

Performance (quantized Qwen3-VL-4B export)

Workload	Before	After
Text decoder layer (prefill)	20.0 s	14.4 s
Vision prefill, t1 (~100k ops)	>2 hours (not finished)	9.4 min
Vision prefill, t8 (~800k ops)	>14 hours (not finished)	2 h 51 min
fake-quantize decompose share	29%	2.5%
Circle O1 rules share	26%	~0%

Co-Authored-By: Cline SR

TICO-DCO-1.0-Signed-off-by: Evgenii Maltsev e.maltsev@samsung.com

[dvsav]

_cache is a public, positional third parameter of the exported function with no type annotation (_cache=None) in an otherwise typed module. The underscore convention on a positional argument is weak protection.

[dvsav]

(Optional)

fuse/linear.py:211-213 still keeps its own private _is_live() calling graph.backward_operators(graph.outputs) per operator — this pass has exactly the O(N²) problem the PR set out to eliminate. Would you consider optimizing this as well (in this PR or in another PR)?

[dvsav]

(Optional) The same O(N)-copy antipattern survives in code the O1 rules call per element

• _utils.tensor_contract() and tensor_object() (_utils.py:196-215) do as_list(graph.subgraph.tensors) on every tensor inspection — once per rule match per candidate.
• CircleGraph.is_constant() (graph.py:223-225) delegates to is_constant_tensor(model, subgraph, idx), which re-lists tensors and buffers per call, and in the zero-byte path walks all operators (graph.py:93-96). It's called per-operator in simplify/identity_ops.py:133,212, fuse/linear.py:747, analysis/purity.py:119-122.
• fuse/linear.py:_linear_state re-lists all operators per operator, and _is_live (above).

[Torrero]

@dvsav Thank you for your comments, I fixed it

[mhs4670go]

LGTM

[dvsav]

👍 LGTM. Thank you!